def conv(model, temp):
model.add(
layers.Conv1D(filters=random.choice(filters),
kernel_size=temp,
activation='relu'))
def build(self):
# input is an image of time_window x 1
input_exg = layers.Input(shape=(self.time_window, 1, self.exog_dim))
input_end = layers.Input(shape=(self.time_window, 1))
op_end = input_end
op_exg = self.build_exog_cnn(input_exg)
print '---> Start of parallel layers'
# build parallel dilated layers
for l in range(self.parallel_layers):
print '--- parallel layer ', l + 1
d = math.pow(2, l)
print 'dialtion rate :', d
op_end = self.res_block(inp=op_end,
num_filters=4,
dilation=[d],
kernel_size=2)
op_exg = self.res_block(inp=op_exg,
num_filters=4,
dilation=[d],
kernel_size=2)
# build combined dilated layers
print '---> Start of combined layers'
comb_op = layers.add([op_exg, op_end])
print comb_op
for l in range(self.parallel_layers,
self.parallel_layers + self.comb_layers):
d = math.pow(2, l)
print 'layer', l + 1, 'dialtion rate :', d
comb_op = self.res_block(inp=comb_op,
num_filters=4,
dilation=[d],
kernel_size=2)
network_op = layers.Conv1D(
filters=1,
kernel_size=1,
dilation_rate=[1],
strides=1,
kernel_regularizer=keras.regularizers.l2(0.01),
padding='valid')(comb_op)
network_op = layers.LeakyReLU()(network_op)
model = models.Model(inputs=[input_exg, input_end],
outputs=[network_op])
model.compile(optimizer=keras.optimizers.Adam(),
loss=keras.losses.MSE,
metrics=[keras.metrics.mae, keras.metrics.mse])
print model.summary()
self.model = model
return
from keras import layers from keras.models import Sequential max_features = 10000 max_len = 500 (x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features) print(len(x_train), len(x_test)) x_train = sequence.pad_sequences(x_train, maxlen=max_len) x_test = sequence.pad_sequences(x_test, maxlen=maxlen) #training and evaluating a simple 1D convnet on the IMDB data from keras.models import Sequential from keras import layers from keras.optimizers import RMSprop model = Sequential() model.add(layers.Embedding(max_features, 128, input_length=max_len)) model.add(layers.Conv1D(32, 7, activation='relu')) model.add(layers.MaxPooling1D(5)) model.add(layers.Conv1D(32, 7, activation='relu')) model.add(layers.GlobalMaxPooling1D()) model.add(layers.Dense(1)) model.summary() model.compile(optimizer=RMSprop(lr=1e-4),loss='binary_crossentropy',metrics=['acc']) history = model.fit(x_train, y_train,epochs=10,batch_size=128,validation_split=0.2) #because 1D convnets process input batches independently, they are not sensitive to the order of timesteps beyond a local scale(size of the convolution windows) #to recorgnize longer-term patterns you can stack many convolution layers and pooling layers, resultingin upper ayers that will see long chunks of the original data #however, this doenst work very well #on temperature dataset - because more recent data points should be interpreted differently from older data points, the convnet fails at producing meaningful results
def ieee_net(x_train, y_train, ddg_train, epoch_best):
row_num, col_num = x_train.shape[1:3]
verbose = 1
batch_size = 64
epochs = epoch_best #[15, 12, 16, 29, 16, 12, 10, 31, 10, 19]
metrics = ('mae', pearson_r, rmse)
def step_decay(epoch):
# drops as progression proceeds, good for sgd
if epoch > 0.9 * epochs:
lr = 0.00001
elif epoch > 0.75 * epochs:
lr = 0.0001
elif epoch > 0.5 * epochs:
lr = 0.001
else:
lr = 0.01
print('lr: %f' % lr)
return lr
lrate = callbacks.LearningRateScheduler(step_decay, verbose=verbose)
# my_callbacks = [
# lrate
# ]
my_callbacks = None
network = models.Sequential()
network.add(
layers.Conv1D(filters=16,
kernel_size=5,
activation='relu',
input_shape=(row_num, col_num)))
network.add(layers.MaxPooling1D(pool_size=2))
network.add(layers.Conv1D(32, 5, activation='relu'))
network.add(layers.MaxPooling1D(pool_size=2))
network.add(layers.Conv1D(64, 3, activation='relu'))
network.add(layers.MaxPooling1D(pool_size=2))
network.add(layers.Flatten())
network.add(layers.Dense(128, activation='relu'))
network.add(layers.Dropout(0.5))
network.add(layers.Dense(16, activation='relu'))
network.add(layers.Dropout(0.3))
network.add(layers.Dense(1))
# print(network.summary())
# rmsp = optimizers.RMSprop(lr=0.0001, decay=0.1)
rmsp = optimizers.RMSprop(lr=0.0001)
network.compile(
optimizer=rmsp, #'rmsprop', # SGD,adam,rmsprop
loss='mse',
metrics=list(metrics)) # mae平均绝对误差（mean absolute error） accuracy
result = network.fit(
x=x_train,
y=ddg_train,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=my_callbacks,
shuffle=True,
)
return network, result.history
def model_creation(input_shape1,
input_shape2,
model_version,
optim,
problem_type=0):
"""
op_sequence : False, by default. If True, then also uses the num_nodes parameter.
num_nodes : The number that denotes that how many values to predict at the output layer.
"""
print("Inside model_creation")
past_inp = L.Input(shape=(input_shape1))
fut_inp = L.Input(shape=(input_shape2))
if (model_version == "M1V1"):
cnn1 = L.Conv1D(filters=32, kernel_size=5)(past_inp)
cnn1 = L.Conv1D(filters=32, kernel_size=2)(cnn1)
cnn2 = L.Conv1D(filters=32, kernel_size=5)(fut_inp)
cnn2 = L.Conv1D(filters=32, kernel_size=2)(cnn2)
lstm_inp = L.Average()([cnn1, cnn2])
lstm_out = L.LSTM(32,
recurrent_dropout=0.2,
return_sequences=True,
bias_initializer='ones')(lstm_inp)
x1 = L.Average()([lstm_out, lstm_inp])
x1 = L.Flatten()(x1)
elif (model_version == "M1V2"):
cnn1 = L.Conv1D(filters=32, kernel_size=5)(past_inp)
cnn1 = L.Conv1D(filters=32, kernel_size=3)(cnn1)
cnn2 = L.Conv1D(filters=32, kernel_size=5)(fut_inp)
cnn2 = L.Conv1D(filters=32, kernel_size=3)(cnn2)
x1 = L.Average()([cnn1, cnn2])
x1 = L.Flatten()(x1)
elif (model_version == "M1V3"):
x1 = L.LSTM(32,
recurrent_dropout=0.2,
return_sequences=True,
bias_initializer='ones')(past_inp)
#x1 = L.LSTM(32, recurrent_dropout=0.2, return_sequences=True, bias_initializer='ones')(x1)
x1 = L.LSTM(32, recurrent_dropout=0.2, bias_initializer='ones')(x1)
elif (model_version == "M2V1"):
cnn1 = L.Conv1D(filters=32, kernel_size=5)(past_inp)
cnn1 = L.Conv1D(filters=32, kernel_size=2)(cnn1)
lstm_out1 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn1)
lstm_out1 = L.Average()([cnn1, lstm_out1])
cnn2 = L.Conv1D(filters=32, kernel_size=5)(fut_inp)
cnn2 = L.Conv1D(filters=32, kernel_size=2)(cnn2)
lstm_out2 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn2)
lstm_out2 = L.Average()([cnn2, lstm_out2])
x1 = L.Average()([lstm_out1, lstm_out2])
x1 = L.Flatten()(x1)
elif (model_version == "M2V2"):
lstm_out1 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(past_inp)
lstm_out2 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(fut_inp)
x1 = L.Average()([lstm_out1, lstm_out2])
x1 = L.Flatten()(x1)
elif (model_version == "M3V1"):
# cnn_inp = L.Concatenate(axis=1)([past_inp,fut_inp])
cnn = L.Conv1D(filters=32, kernel_size=5)(past_inp)
cnn = L.Conv1D(filters=32, kernel_size=2)(cnn)
layer = L.Lambda(lambda cnn: K.reverse(cnn, axes=1))
cnn2 = layer(cnn)
lstm1 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn)
lstm2 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn2)
lstm = L.Average()([lstm1, lstm2])
x1 = L.Average()([cnn, lstm])
x1 = L.Flatten()(x1)
elif (model_version == "M3V2"):
cnn_inp = L.Concatenate(axis=1)([past_inp, fut_inp])
cnn = L.Conv1D(filters=32, kernel_size=5)(cnn_inp)
cnn = L.Conv1D(filters=32, kernel_size=2)(cnn)
layer = L.Lambda(lambda cnn: K.reverse(cnn, axes=1))
cnn2 = layer(cnn)
lstm1 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn)
lstm2 = L.LSTM(32,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones')(cnn2)
lstm = L.Average()([lstm1, lstm2])
x1 = L.Average()([cnn, lstm])
x1 = L.Flatten()(x1)
elif (model_version == "M4"):
cnn1 = L.Conv1D(filters=32, kernel_size=5)(past_inp)
cnn1 = L.Conv1D(filters=32, kernel_size=2)(cnn1)
cnn2 = L.Conv1D(filters=32, kernel_size=5)(fut_inp)
cnn2 = L.Conv1D(filters=32, kernel_size=2)(cnn2)
lstm1 = L.Bidirectional(
L.LSTM(16,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones'))(past_inp)
lstm2 = L.Bidirectional(
L.LSTM(16,
recurrent_dropout=0.3,
return_sequences=True,
bias_initializer='ones'))(fut_inp)
out1 = L.Concatenate(axis=1)([cnn1, lstm1])
out2 = L.Concatenate(axis=1)([cnn2, lstm2])
# out1 = L.Average()([cnn1,cnn2])
# ousst2 = L.Average()([lstm1,lstm2])
# x1 = L.Concatenate(axis=1)([out1,out2])
x1 = L.Average()([out1, out2])
x1 = L.Flatten()(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
x1 = L.Dense(256)(x1)
x1 = L.advanced_activations.LeakyReLU(0.2)(x1)
#Classification Part.
if (problem_type == 1):
main_out = L.Dense(2, activation='softmax')(x1)
model = M.Model(inputs=[past_inp, fut_inp],
outputs=[main_out],
name=model_version)
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
#Regression Part.
else:
x1 = L.Dense(1)(x1)
main_out = L.advanced_activations.LeakyReLU(0.2)(x1)
model = M.Model(inputs=[past_inp, fut_inp],
outputs=[main_out],
name=model_version)
model.compile(optimizer=optim,
loss=tf.losses.huber_loss,
metrics=['mae', 'mse', rmse])
model.summary()
return model
def au_Exp():
now = datetime.datetime.now()
now_s = now.strftime("%Y-%m-%d-%H-%M-%S")
config = tf.compat.v1.ConfigProto(gpu_options=tf.compat.v1.GPUOptions(
allow_growth=True))
sess = tf.compat.v1.Session(config=config)
## 准备几个参数，用于后续的自动化
epochs_au = 50
batch_size_au = 1
jihuo = 'tanh'
callback_list_test = [
keras.callbacks.ModelCheckpoint(
filepath=now_s + '.h5', ##文件路径 存在当前路径下吧 还好找
monitor='val_loss', ## 监控指标
save_best_only=True ## 保持最佳模型
)
]
# dataspecimen,dataprop = gM.getOdataMilk()
# test_data,test_lable,train_data,train_lable = sF.getTestDataMilk(dataspecimen,dataprop)
test_data, test_lable, train_data, train_lable = mD.getOdatebmilkDep()
model = models.Sequential()
model.add(layers.Conv1D(64, 7, activation=jihuo,
input_shape=(1557, 1))) #输入形状依据样品进行变换，此处牛奶为1577
model.add(layers.MaxPooling1D(2))
model.add(layers.Conv1D(32, 7, activation=jihuo))
model.add(layers.MaxPooling1D(2))
model.add(layers.Conv1D(32, 7, activation=jihuo))
model.add(layers.MaxPooling1D(2))
model.add(layers.Conv1D(16, 7, activation=jihuo))
# model.add(layers.GlobalMaxPooling1D()) ## 实际效果极差1
model.add(layers.Flatten())
model.add(layers.Dense(16))
model.add(layers.Dense(8))
# model.add(layers.Dense(4))
# model.add(layers.Dense(2))
model.add(layers.Dense(1))
model.summary()
model.compile(optimizer=RMSprop(), loss='mse')
history = model.fit(train_data,
train_lable,
epochs=epochs_au,
batch_size=batch_size_au,
validation_data=(test_data, test_lable),
callbacks=callback_list_test)
sF.drawLoss(history) ## 绘制当前的验证曲线
model = load_model(now_s + '.h5')
result_trian = model.predict(train_data)
result_predict = model.predict(test_data)
rmsec = sF.calculate_RMSE(result_trian, train_lable) ## 训练集上的RMSE
rmsep = sF.calculate_RMSE(result_predict, test_lable) ## 测试集上的RMSE
r_2_t = sF.calculate_R21(result_trian, train_lable) ## 训练集上的R_2
r_2_p = sF.calculate_R21(result_predict, test_lable) ## 测试集上得R_2
# print("Root Mean Square Error of Calibrationh is : %g"%(rmsec))
# print("训练集上得决定系数：%f"%(r_2_t))
# print("Root Mean Square Error of Prediction is : %g"%(rmsep))
# print("测试集上得决定系数：%f"%(r_2_p))
###### 下面的代码用于自动记录实验数据
write_data = [(now_s, epochs_au, batch_size_au, rmsec, r_2_t, rmsep, r_2_p)
] #需要新写入的数据
sF.write_To_Csv(write_data)
return rmsep, r_2_p
row_data = data.iloc[i:i+1,0].values[0]
row_data = preprocessing(row_data)
row_data = cut_word(row_data)
input_text.append(row_data)
print(i)
tf_data = vector.fit_transform(input_text)
# model
from keras import layers , Input
from keras.models import Model
vocabulary_size = len(data)
num_income_groups = 10
posts_input = Input(shape = (None,),name='posts')
embedded_posts = layers.Embedding(256,vocabulary_size)(posts_input)
x = layers.Conv1D(128,5,activation = 'relu')(embedded_posts)
x = layers.MaxPooling1D(5)(x)
x = layers.Conv1D(256,5,activation = 'relu')(x)
x = layers.Conv1D(256,5,activation = 'relu')(x)
x = layers.MaxPooling1D(5)(x)
x = layers.Conv1D(256,5,activation = 'relu')(x)
x = layers.Conv1D(256,5,activation = 'relu')(x)
x = layers.Dense(128,activation = 'relu')(x)
toxic_prediction = layers.Dense(1,activation = 'sigmoid',name = 'toxic')(x)
severe_toxic_prediction = layers.Dense(1,activation = 'sigmoid',name = 'severe_toxic')(x)
obscene_prediction = layers.Dense(1,activation = 'sigmoid',name = 'obscene')(x)
threat_prediction = layers.Dense(1,activation = 'sigmoid',name = 'threat')(x)
insult_prediction = layers.Dense(1,activation = 'sigmoid',name = 'insult')(x)
identity_hate = layers.Dense(1,activation = 'sigmoid',name = 'identity_hate')(x)
model = Sequential()
#model.add(layers.Embedding(input_dim=vocab_size,
# output_dim=embedding_dim,
# embeddings_initializer = glorot_uniform(seed=42),
# input_length=maxlen,
# trainable=True))
model.add(
layers.Embedding(vocab_size,
embedding_dim,
weights=[embedding_matrix],
input_length=maxlen,
trainable=False))
model.add(layers.Conv1D(70, 2, activation='relu'))
model.add(layers.MaxPooling1D(3))
model.add(layers.Conv1D(100, 3, activation='relu'))
model.add(layers.GlobalMaxPool1D())
model.add(layers.Dense(60, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
from keras.optimizers import SGD
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy',
optimizer='sgd',
metrics=[recall, precision, f1, 'accuracy'])
model.summary()
history = model.fit(X_train,
def build_model(self):
"""
Функция, задающая архитектуру нейронной сети
"""
# symbol_inputs: array, 1D-массив длины m
symbol_inputs = kl.Input(shape=(None,), dtype='uint8', name="symbol_inputs")
# symbol_embeddings: array, 2D-массив размера m*self.symbols_number
if self.use_embeddings:
symbol_embeddings = kl.Embedding(self.symbols_number_, self.embeddings_size,
name="symbol_embeddings")(symbol_inputs)
else:
symbol_embeddings = kl.Lambda(kb.one_hot, output_shape=(None, self.symbols_number_),
arguments={"num_classes": self.symbols_number_},
name="symbol_embeddings")(symbol_inputs)
inputs = [symbol_inputs]
if self.to_memorize_morphemes:
# context_inputs: array, 2D-массив размера m*15
context_inputs = kl.Input(shape=(None, self.memory_dim), dtype='float32', name="context_inputs")
inputs.append(context_inputs)
if self.context_dropout > 0.0:
context_inputs = kl.Dropout(self.context_dropout)(context_inputs)
# представление контекста подклеивается к представлению символа
symbol_embeddings = kl.Concatenate()([symbol_embeddings, context_inputs])
conv_inputs = symbol_embeddings
conv_outputs = []
for window_size, curr_filters_numbers in zip(self.window_size, self.filters_number):
# свёрточный слой отдельно для каждой ширины окна
curr_conv_input = conv_inputs
for j, filters_number in enumerate(curr_filters_numbers[:-1]):
# все слои свёртки, кроме финального (после них возможен dropout)
curr_conv_input = kl.Conv1D(filters_number, window_size,
activation="relu", padding="same")(curr_conv_input)
if self.dropout > 0.0:
# между однотипными слоями рекомендуется вставить dropout
curr_conv_input = kl.Dropout(self.dropout)(curr_conv_input)
if not self.use_lstm:
curr_conv_output = kl.Conv1D(curr_filters_numbers[-1], window_size,
activation="relu", padding="same")(curr_conv_input)
else:
curr_conv_output = curr_conv_input
conv_outputs.append(curr_conv_output)
# соединяем выходы всех свёрточных слоёв в один вектор
if len(conv_outputs) == 1:
conv_output = conv_outputs[0]
else:
conv_output = kl.Concatenate(name="conv_output")(conv_outputs)
if self.use_lstm:
conv_output = kl.Bidirectional(
kl.LSTM(self.lstm_units, return_sequences=True))(conv_output)
if self.dense_output_units:
pre_last_output = kl.TimeDistributed(
kl.Dense(self.dense_output_units, activation="relu"),
name="pre_output")(conv_output)
else:
pre_last_output = conv_output
# финальный слой с softmax-активацией, чтобы получить распределение вероятностей
output = kl.TimeDistributed(
kl.Dense(self.target_symbols_number_, activation="softmax"), name="output")(pre_last_output)
model = Model(inputs, [output])
model.compile(optimizer=Adam(clipnorm=5.0),
loss="categorical_crossentropy", metrics=["accuracy"])
return model
# 计算val_steps，这是计算需要从val_gen中抽取多少次
val_steps = (valnum - lookback) // batch_size
# 查看训练集需要抽取的次数
train_steps = trainnum // batch_size
from keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor="val_loss", patience=30)
#创建模型
model1 = models.Sequential()
model1.add(
layers.Dense(512,
activation="relu",
input_shape=(lookback // step, data.shape[-1])))
model1.add(layers.Conv1D(filters=1024, kernel_size=5, activation="relu"))
model1.add(layers.MaxPooling1D(5))
model1.add(
layers.Bidirectional(layers.LSTM(32, activation="relu", dropout=0.5)))
model1.add(layers.Dense(8, activation="softmax"))
model1.summary()
model1.compile(optimizer=optimizers.RMSprop(),
loss="categorical_crossentropy",
metrics=["acc"])
history1 = model1.fit_generator(train_gen,
steps_per_epoch=train_steps,
epochs=200,
validation_data=val_gen,
validation_steps=val_steps,
callbacks=[early_stopping])
#pred_win = 72
#lookback = 144
dense_units = pred_win
cn = 1
for batch_size in batch_size_list:
for lr in lr_list:
print('batch_size=', batch_size, 'lr = ', lr, 'cn = ', cn, 'epochs =',
epochs, 'delay=', delay, 'lookback=', lookback, 'pred_win=',
pred_win, 'components =', n_comp)
model = models.Sequential()
#model.add(layers.Embedding(max_features, 128, input_length=max_len))
model.add(
layers.Conv1D(32,
7,
activation='relu',
input_shape=(x_seq_train.shape[1], n_features)))
#model.add(layers.MaxPooling1D(5))
#model.add(layers.Conv1D(256, 7, activation='relu'))
#model.add(layers.GlobalMaxPooling1D())
model.add(MaxPooling1D(pool_size=7))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(layers.Dense(y_seq_train.shape[1]))
model.summary()
model.compile(optimizer=Adam(lr=lr, clipnorm=cn), loss='mse')
# fit model
history = model.fit(x_seq_train,
y_seq_train,
def build_model(self, X_train):
# input layers
main_input = layers.Input(shape=(len(self.main_input_features), ),
dtype="float32",
name="main_input")
lag_player_stats_input = layers.Input(
shape=X_train["lag_player_stats_input"].shape[1:],
dtype="float32",
name="lag_player_stats_input",
)
lag_opp_stats_input = layers.Input(
shape=X_train["lag_opp_stats_input"].shape[1:],
dtype="float32",
name="lag_opp_stats_input",
)
lag_team_total_stats_input = layers.Input(
shape=X_train["lag_team_total_stats_input"].shape[1:],
dtype="float32",
name="lag_team_total_stats_input",
)
lag_opp_total_stats_input = layers.Input(
shape=X_train["lag_opp_total_stats_input"].shape[1:],
dtype="float32",
name="lag_opp_total_stats_input",
)
# convolutional layers
conv_lag_player_stats = layers.Conv1D(
filters=self.conv_filters,
kernel_size=self.conv_kernel_size,
activation="relu",
)(lag_player_stats_input)
conv_lag_player_stats = layers.MaxPooling1D(
pool_size=self.pool_size)(conv_lag_player_stats)
conv_lag_player_stats = layers.Flatten()(conv_lag_player_stats)
conv_lag_opp_stats = layers.Conv1D(
filters=self.conv_filters,
kernel_size=self.conv_kernel_size,
activation="relu",
)(lag_opp_stats_input)
conv_lag_opp_stats = layers.MaxPooling1D(
pool_size=self.pool_size)(conv_lag_opp_stats)
conv_lag_opp_stats = layers.Flatten()(conv_lag_opp_stats)
conv_lag_team_total_stats = layers.Conv1D(
filters=self.conv_filters,
kernel_size=self.conv_kernel_size,
activation="relu",
)(lag_team_total_stats_input)
conv_lag_team_total_stats = layers.MaxPooling1D(
pool_size=self.pool_size)(conv_lag_team_total_stats)
conv_lag_team_total_stats = layers.Flatten()(conv_lag_team_total_stats)
conv_lag_opp_total_stats = layers.Conv1D(
filters=self.conv_filters,
kernel_size=self.conv_kernel_size,
activation="relu",
)(lag_opp_total_stats_input)
conv_lag_opp_total_stats = layers.MaxPooling1D(
pool_size=self.pool_size)(conv_lag_opp_total_stats)
conv_lag_opp_total_stats = layers.Flatten()(conv_lag_opp_total_stats)
# main layers
concat_all = layers.concatenate([
main_input,
conv_lag_player_stats,
conv_lag_opp_stats,
conv_lag_team_total_stats,
conv_lag_opp_total_stats,
])
concat_all = layers.Dense(self.normal_layer_size,
activation="relu",
bias_initializer="zeros")(concat_all)
concat_all = layers.Dropout(0.2)(concat_all)
# output layers
main_output = layers.Dense(
self.upper - self.lower + 1,
activation="sigmoid",
kernel_constraint=constraints.unit_norm(axis=0),
name="main_output",
)(concat_all)
# define model
input_layers = [
main_input,
lag_player_stats_input,
lag_opp_stats_input,
lag_team_total_stats_input,
lag_opp_total_stats_input,
]
output_layers = [main_output]
output_losses = {"main_output": "binary_crossentropy"}
output_loss_weights = {"main_output": 1}
opt = optimizers.Adam(lr=self.learning_rate)
model = models.Model(inputs=input_layers, outputs=output_layers)
model.compile(optimizer=opt,
loss=output_losses,
loss_weights=output_loss_weights)
return model
def getTextNet():
words_input = kl.Input(shape=(max_wlen, ), name='words_input')
padding_masks = kl.Input(shape=(max_wlen, 1), name='padding_masks')
x2 = kl.Embedding(vocab_size + 1,
embedding_size,
mask_zero=False,
name='w2v_emb')(words_input)
xk3 = kl.Conv1D(filters=324, kernel_size=3, strides=1, padding='same')(x2)
xk3 = kl.ELU(alpha=elu_alpha)(xk3)
xk5 = kl.Conv1D(filters=324, kernel_size=5, strides=1, padding='same')(x2)
xk5 = kl.ELU(alpha=elu_alpha)(xk5)
xk7 = kl.Conv1D(filters=324, kernel_size=7, strides=1, padding='same')(x2)
xk7 = kl.ELU(alpha=elu_alpha)(xk7)
xk3d2 = kl.Conv1D(filters=324,
kernel_size=3,
strides=1,
dilation_rate=2,
padding='same')(x2)
xk3d2 = kl.ELU(alpha=elu_alpha)(xk3d2)
xk5d2 = kl.Conv1D(filters=324,
kernel_size=5,
strides=1,
dilation_rate=2,
padding='same')(x2)
xk5d2 = kl.ELU(alpha=elu_alpha)(xk5d2)
xk7d2 = kl.Conv1D(filters=324,
kernel_size=7,
strides=1,
dilation_rate=2,
padding='same')(x2)
xk7d2 = kl.ELU(alpha=elu_alpha)(xk7d2)
x2 = kl.Concatenate()([xk3, xk5, xk7, xk3d2, xk5d2, xk7d2])
# x2 = kl.BatchNormalization()(x2)
# x2 = kl.ELU(alpha=elu_alpha)(x2)
x2 = kl.Conv1D(filters=100, kernel_size=1, strides=1, padding='same')(x2)
x2 = kl.BatchNormalization()(x2)
x2 = kl.ELU(alpha=elu_alpha)(x2)
# print('x2.shape:',x2.shape)
x2 = kl.Dropout(dropout_rate)(x2)
sa_out_x2_1, s_x2_1 = SelfAttention(ch=int(x2.shape[-1]),
name='sa_2_1')([x2, padding_masks])
sa_out_x2_2, s_x2_2 = SelfAttention(ch=int(x2.shape[-1]),
name='sa_2_2')([x2, padding_masks])
sa_out_x2_3, s_x2_3 = SelfAttention(ch=int(x2.shape[-1]),
name='sa_2_3')([x2, padding_masks])
sa_out_x2_4, s_x2_4 = SelfAttention(ch=int(x2.shape[-1]),
name='sa_2_4')([x2, padding_masks])
# print(sa_out_x2_4)
x3 = kl.Concatenate(name='concat_sa_2')(
[sa_out_x2_1, sa_out_x2_2, sa_out_x2_3, sa_out_x2_4])
# x3 = kl.ELU(alpha=elu_alpha,name='act_concat_sa_2')(x3)
x3comb, x3_g1, x3_g2 = ResidualCombine1D(ch_in=int(x3.shape[-1]),
ch_out=100)([x2, x3])
x3comb = kl.BatchNormalization()(x3comb)
# x3comb = kl.ELU(alpha=elu_alpha)(x3comb)
x3comb = kl.Conv1D(filters=100, kernel_size=1, strides=1,
padding='same')(x3comb)
x3comb = kl.BatchNormalization()(x3comb)
x3comb = kl.ELU()(x3comb)
x3comb = kl.Dropout(dropout_rate)(x3comb)
sa_out_x3_1, s_x3_1 = SelfAttention(ch=int(x3comb.shape[-1]),
name='sa_3_1')([x3comb, padding_masks])
sa_out_x3_2, s_x3_2 = SelfAttention(ch=int(x3comb.shape[-1]),
name='sa_3_2')([x3comb, padding_masks])
sa_out_x3_3, s_x3_3 = SelfAttention(ch=int(x3comb.shape[-1]),
name='sa_3_3')([x3comb, padding_masks])
sa_out_x3_4, s_x3_4 = SelfAttention(ch=int(x3comb.shape[-1]),
name='sa_3_4')([x3comb, padding_masks])
x4 = kl.Concatenate(name='concat_sa_3')(
[sa_out_x3_1, sa_out_x3_2, sa_out_x3_3, sa_out_x3_4])
# x4 = kl.ELU(alpha=elu_alpha,name='act_concat_sa_3')(x4)
x4comb, x4_g1, x4_g2 = ResidualCombine1D(ch_in=int(x4.shape[-1]),
ch_out=100)([x3comb, x4])
x4comb = kl.BatchNormalization()(x4comb)
# x4comb = kl.ELU(alpha=elu_alpha)(x4comb)
x4comb = kl.Conv1D(filters=100, kernel_size=1, strides=1,
padding='same')(x4comb)
x4comb = kl.BatchNormalization()(x4comb)
x4comb = kl.ELU(alpha=elu_alpha)(x4comb)
# x4comb = kl.Dropout(dropout_rate)(x4comb)
# sa_out_x4_1,s_x4_1 = SelfAttention(ch=int(x4comb.shape[-1]),name='sa_4_1')([x4comb,padding_masks])
# sa_out_x4_2,s_x4_2 = SelfAttention(ch=int(x4comb.shape[-1]),name='sa_4_2')([x4comb,padding_masks])
# sa_out_x4_3,s_x4_3 = SelfAttention(ch=int(x4comb.shape[-1]),name='sa_4_3')([x4comb,padding_masks])
# sa_out_x4_4,s_x4_4 = SelfAttention(ch=int(x4comb.shape[-1]),name='sa_4_4')([x4comb,padding_masks])
# x5 = kl.Concatenate(name='concat_sa_4')([sa_out_x4_1,sa_out_x4_2,sa_out_x4_3,sa_out_x4_4])
# x5 = kl.ELU(alpha=elu_alpha,name='act_concat_sa_4')(x5)
# x5comb,x5_g1,x5_g2 = ResidualCombine1D(ch_in=int(x5.shape[-1]),ch_out=256)([x4comb,x5])
# x5comb = kl.BatchNormalization()(x5comb)
# x5comb = kl.ELU(alpha=elu_alpha)(x5comb)
# x5comb = kl.Conv1D(filters=256,kernel_size=1,strides=1,padding='same')(x5comb)
# x5comb = kl.BatchNormalization()(x5comb)
# x5comb = kl.ELU(alpha=elu_alpha)(x5comb)
return Model([words_input, padding_masks], x4comb, name='textModel')
axis=1),
axis=2)
train_y = np.concatenate((train_y, n_y))
train_X = np.vstack((train_X, n_X))
train_X, train_y = shuffle(train_X, train_y, random_state=0)
#Resample
if len(classes) == 2:
uniques, counts = np.unique(train_y, return_counts=True)
print(counts)
if len(classes) != 2:
train_y = keras.utils.to_categorical(train_y)
model = keras.Sequential()
model.add(layers.Conv1D(64, 3, input_shape=(20, 66), activation="relu"))
model.add(layers.Conv1D(64, 3, activation="relu"))
model.add(layers.MaxPooling1D(3))
model.add(layers.Conv1D(128, 3, activation="relu"))
model.add(layers.GlobalAveragePooling1D())
model.add(layers.Dropout(rate=0.5))
adam = keras.optimizers.Adam(lr=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
decay=0.0,
amsgrad=True)
if len(classes) == 2:
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
kfold = StratifiedKFold(n_splits=5, shuffle=True, random_state=2018)
cvscores_cnn = []
dataset_blend_train_cnn = np.zeros((X_train.shape[0], 1))
preds_test_blend_cnn = np.zeros((X_test.shape[0], 5))
for i, (train_idx, test_idx) in enumerate(kfold.split(X_train, y)):
X_t, y_t, X_te, y_te = X_train[train_idx], y[train_idx], X_train[
test_idx], y[test_idx]
cnn_model0 = Sequential()
cnn_model0.add(
Embedding(WORD_IDX_LEN + 1,
EMBEDDING_DIM,
weights=[embedding_matrix],
input_length=MAX_LEN,
trainable=False))
cnn_model0.add(layers.Conv1D(128, 5, activation='relu'))
cnn_model0.add(layers.GlobalMaxPooling1D())
cnn_model0.add(layers.Dense(512, activation='relu'))
cnn_model0.add(layers.Dense(512, activation='relu'))
cnn_model0.add(Dropout(0.1))
cnn_model0.add(layers.Dense(512, activation='relu'))
cnn_model0.add(Dropout(0.1))
cnn_model0.add(layers.Dense(512, activation='relu'))
cnn_model0.add(Dropout(0.1))
cnn_model0.add(layers.Dense(1))
cnn_model0.compile(optimizer='rmsprop', loss='mse', metrics=[escore])
hist_cnn = cnn_model0.fit(X_t,
y_t,
epochs=10,
batch_size=64,
validation_data=(X_te, y_te),
############################################################################
results = {}
from keras.models import Sequential
from keras import layers
# Neural network
model = Sequential()
model.add(
layers.Embedding(vocab_size,
embedding_dim,
weights=[W2V_matrix],
input_length=maxlen,
trainable=True))
model.add(layers.Conv1D(128, 5, activation='relu'))
model.add(layers.GlobalMaxPooling1D())
model.add(layers.Dense(10, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
model.summary()
model.fit(X_train,
y_train,
epochs=3,
verbose=True,
validation_data=(X_validation, y_validation),
batch_size=10)
results['word2vec'] = model
def build_model(self, embedding_layer):
"""
Builds the keras model corresponding to specified type of architecture
:param embedding_layer: preloaded embedding layer for keras model
:return: keras model
"""
model = Sequential()
model.add(embedding_layer)
#3-layered bidirectional LSTM
if self.arch_name == "NN":
model.add(layers.Flatten())
model.add(layers.Dense(1024, activation='relu'))
model.add(layers.Dense(2, activation='softmax'))
elif self.arch_name == "BiLSTM":
model.add(
layers.Bidirectional(
layers.LSTM(100, return_sequences=True, dropout=0.1)))
model.add(
layers.Bidirectional(layers.LSTM(100, return_sequences=True)))
model.add(layers.Bidirectional(layers.LSTM(100)))
model.add(layers.Dense(50, activation='relu'))
model.add(layers.Dense(2, activation='softmax'))
#two-layered forward LSTM
elif self.arch_name == "LSTM":
model.add(layers.LSTM(200, return_sequences=True))
model.add(layers.LSTM(100, return_sequences=False))
model.add(layers.Dense(1000, activation='relu'))
model.add(layers.Dense(2, activation='softmax'))
#CNN model with attention layer
elif self.arch_name == "CNN":
model.add(layers.Conv1D(100, 7, activation='relu', padding='same'))
model.add(layers.MaxPooling1D(2))
model.add(layers.Conv1D(64, 7, activation='relu', padding='same'))
model.add(Attention1())
model.add(layers.Dense(2, activation='sigmoid'))
#single-layered bidirectional GRU with attention layer
elif self.arch_name == "GRU":
model.add(
layers.Bidirectional(
layers.GRU(200, return_sequences=True, dropout=0.1)))
model.add(Attention2(self.ln))
model.add(layers.Dense(200, activation='relu'))
model.add(layers.Dense(2, activation='softmax'))
#Stacked CNN and LSTM with attention layer
elif self.arch_name == "CNN_LSTM":
model.add(layers.Conv1D(128, 3, activation='relu', padding='valid')
) # filters: 100, kernel_size = 7 -> output = 100
model.add(layers.MaxPooling1D(2))
model.add(
layers.Bidirectional(layers.LSTM(128, return_sequences=True)))
model.add(Attention1())
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(2, activation='softmax'))
else:
raise NotImplementedError(
"Specified architecture is not implemented:", self.arch_name)
return model
def CapsNet_nogradientstop(input_shape, n_class, routings):
x = layers.Input(shape=input_shape)
conv1 = layers.Conv1D(filters=200,
kernel_size=1,
strides=1,
padding='valid',
kernel_initializer='he_normal',
activation='relu',
name='conv1')(x)
# conv1=BatchNormalization()(conv1)
conv1 = Dropout(0.7)(conv1)
conv2 = layers.Conv1D(filters=200,
kernel_size=9,
strides=1,
padding='valid',
kernel_initializer='he_normal',
activation='relu',
name='conv2')(conv1)
# conv1=BatchNormalization()(conv1)
conv2 = Dropout(0.7)(conv2) # 0.75 valx loss has 0.1278!
primarycaps = PrimaryCap(conv2,
dim_capsule=8,
n_channels=60,
kernel_size=20,
kernel_initializer='he_normal',
strides=1,
padding='valid',
dropout=0.2)
dim_capsule_dim2 = 10
# Capsule layer. Routing algorithm works here.
digitcaps_c = CapsuleLayer_nogradient_stop(num_capsule=n_class,
dim_capsule=dim_capsule_dim2,
num_routing=routings,
name='digitcaps',
kernel_initializer='he_normal',
dropout=0.1)(primarycaps)
# digitcaps_c = CapsuleLayer(num_capsule=n_class, dim_capsule=dim_capsule_dim2, num_routing=routings,name='digitcaps',kernel_initializer='he_normal')(primarycaps)
digitcaps = Extract_outputs(dim_capsule_dim2)(digitcaps_c)
weight_c = Extract_weight_c(dim_capsule_dim2)(digitcaps_c)
out_caps = Length(name='capsnet')(digitcaps)
# Decoder network.
y = layers.Input(shape=(n_class, ))
masked_by_y = Mask()(
[digitcaps, y]
) # The true label is used to mask the output of capsule layer. For training
masked = Mask(
)(digitcaps) # Mask using the capsule with maximal length. For prediction
# Shared Decoder model in training and prediction
decoder = Sequential(name='decoder')
decoder.add(
layers.Dense(512,
activation='relu',
input_dim=dim_capsule_dim2 * n_class))
decoder.add(layers.Dense(1024, activation='relu'))
decoder.add(layers.Dense(np.prod(input_shape), activation='sigmoid'))
decoder.add(layers.Reshape(target_shape=input_shape, name='out_recon'))
# Models for training and evaluation (prediction)
train_model = Model([x, y], [out_caps, decoder(masked_by_y)])
eval_model = Model(x, [out_caps, decoder(masked)])
weight_c_model = Model(x, weight_c)
# manipulate model
noise = layers.Input(shape=(n_class, dim_capsule_dim2))
noised_digitcaps = layers.Add()([digitcaps, noise])
masked_noised_y = Mask()([noised_digitcaps, y])
manipulate_model = Model([x, y, noise], decoder(masked_noised_y))
return train_model, eval_model, manipulate_model, weight_c_model
X_test = np.reshape(
X_test,
(np.size(X_test, 0), 1, np.size(X_test, 1), np.size(X_test, 2)))
X_val = np.reshape(
X_val, (np.size(X_val, 0), 1, np.size(X_val, 1), np.size(X_val, 2)))
# Choose hyperparameters
no_epochs = 25
batch_size = 32
learning_rate = 0.0100
dropout_rate = 0.05
# Design the Network
model = Sequential()
model.add(
layers.Conv1D(64, 6, activation='relu', input_shape=X_train[0].shape)
) # Input shape is VERY fiddly. May need to try different things.
model.add(Dropout(dropout_rate))
model.add(layers.Conv1D(128, 6, activation='relu'))
model.add(layers.GlobalMaxPooling1D())
model.add(Dense(numberTargets, activation='softmax'))
print(model.summary())
"""
# Design the Network
model = Sequential()
model.add(layers.SeparableConv2D(32, (1, 2), input_shape=X_train[0].shape, activation='relu'))
model.add(Dropout(dropout_rate))
#model.add(layers.BatchNormalization()) # BatchNormalization and dropout work poorly together, though - Ioffe & Szegedy 2015; Li et al. 2018
model.add(layers.SeparableConv2D(64, (1, 3), activation='sigmoid'))
model.add(layers.SeparableConv2D(128, (1, 4), activation='sigmoid'))
model.add(Dropout(dropout_rate))
print(Counter(target_val))
print(daily_train.shape, daily_val.shape, daily_test.shape, fif_train.shape,
fif_val.shape, fif_test.shape, target_train.shape, target_val.shape,
target_test.shape)
# print(daily_tradaily_train.shape,in,daily_test,fif_train,fif_test,target_train,target_test)
# np.set_printoptions(threshold='nan')
# print(daily_train.shape,daily_test.shape,fif_train.shape,fif_test.shape,target_train.shape,target_test.shape)
# print(daily_train,daily_test,fif_train,fif_test,target_train,target_test)
##### 一、模型搭建
# 15分钟频输入训练(!!!卷积滤镜行列先后)
fif_min_input = Input(shape=(16, 5), dtype='float32', name='fif_min_input')
# fif_min_input=(8,16,4,1)
Conv1D_fif = layers.Conv1D(16, 1, strides=1)(fif_min_input)
LSTM_fif = layers.LSTM(100)(Conv1D_fif)
# 日频输入训练
daily_input = Input(shape=(20, 5), dtype='float32', name='daily_input')
# daily_input=(8,16,4,1)
Conv1D_daily = layers.Conv1D(16, 1, strides=1)(daily_input)
LSTM_daily = layers.LSTM(100)(Conv1D_daily)
# 15分钟频训练结果和日频训练结果合并
concatenated = layers.concatenate([LSTM_fif, LSTM_daily],
axis=-1) # axis=-1按照最后一个轴粘合
alloy = layers.Dense(20, activation='relu')(concatenated) #将粘合结果再接一个全连接层
dropout = layers.Dropout(0.2)(alloy)
output = layers.Dense(1, activation='sigmoid')(dropout)
model = Model([fif_min_input, daily_input], output) #八股文：将输入和输出圈起来
vect_X_train, axis=1
) # final dataframe for training fold of cross validation
X_test_dtm = np.concatenate(
vect_X_test, axis=1
) # final dataframe for testing fold of cross validation
X_train_dtm = np.expand_dims(
X_train_dtm, axis=2
) # add 1 dimension to fit to the convolutionnal layer
X_test_dtm = np.expand_dims(
X_test_dtm, axis=2
) # add 1 dimension to fit to the convolutionnal layer
### MODEL ###
model = models.Sequential()
model.add(
layers.Conv1D(filters=128,
kernel_size=(4),
activation='relu',
input_shape=X_train_dtm[0].shape))
model.add(layers.GlobalMaxPooling1D())
model.add(layers.Dropout(rate=.4))
model.add(
layers.Dense(units=len(target_names),
activation='softmax'))
model.compile(optimizer="adam",
loss="sparse_categorical_crossentropy",
metrics=['accuracy'])
model.fit(X_train_dtm,
y_train,
epochs=epochs,
batch_size=20,
class_weight=class_weight,
validation_data=(X_test_dtm, y_test))
for train, test in kfold.split(x_train, y_train):
tokenizer = Tokenizer()
tokenizer.fit_on_texts(x_train)
encoded_train = tokenizer.texts_to_sequences(x_train[train])
encoded_test = tokenizer.texts_to_sequences(x_train[test])
Xtrain = pad_sequences(encoded_train, maxlen=max_length, padding='post')
Xtest = pad_sequences(encoded_test, maxlen=max_length, padding='post')
vocab_size = len(tokenizer.word_index) + 1
# CNN model
model = keras.Sequential()
# words will be embedded into 100-long vector (output)
# turn positive integers (indexes) into dense vectors of fixed size
model.add(layers.Embedding(vocab_size, 100, input_length=max_length))
model.add(layers.Conv1D(filters=32, kernel_size=8, activation='relu'))
model.add(layers.MaxPooling1D(pool_size=2))
model.add(layers.Flatten())
model.add(layers.Dense(10, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
print(model.summary())
Xtrain = Xtrain.astype(np.float32)
Xtest = Xtest.astype(np.float32)
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print(
'-------------------------------------------------------------------------'
)
print(f'Training for fold {fold_no} ...')
print(train_data_multi.shape)
print(test_data_multi.shape)
print(val_data_multi.shape)
# # Building and Testing the Neural Networks
# In[29]:
# Model 1: Initial model on the simple data set
opt = keras.optimizers.adam(lr=0.01)
CNN_above_in_1 = models.Sequential()
CNN_above_in_1.add(
layers.Conv1D(filters=6,
kernel_size=300,
activation='relu',
input_shape=(7500, 1)))
CNN_above_in_1.add(layers.Conv1D(filters=4, kernel_size=4, activation='relu'))
CNN_above_in_1.add(layers.Dropout(0.5))
CNN_above_in_1.add(layers.MaxPooling1D(pool_size=2))
CNN_above_in_1.add(layers.Flatten())
CNN_above_in_1.add(layers.Dense(44, activation='relu'))
CNN_above_in_1.add(layers.Dense(22, activation='softmax'))
CNN_above_in_1.compile(loss='binary_crossentropy',
optimizer=opt,
metrics=['accuracy'])
# In[26]:
'''
#To save the model weights to a hard drive
# 计算val_steps，这是计算需要从val_gen中抽取多少次
val_steps = (valnum - lookback) // batch_size
# 查看训练集需要抽取的次数
train_steps = trainnum // batch_size
from keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor="val_loss", patience=30)
#创建模型
model1 = models.Sequential()
model1.add(
layers.Dense(512,
activation="relu",
input_shape=(lookback // step, data.shape[-1])))
model1.add(layers.Conv1D(filters=32, kernel_size=3, activation="relu"))
model1.add(layers.MaxPooling1D(2))
model1.add(layers.Dropout(0.5))
model1.add(layers.Conv1D(filters=32, kernel_size=3, activation="relu"))
model1.add(layers.MaxPooling1D(2))
model1.add(layers.Dropout(0.5))
model1.add(layers.Conv1D(filters=32, kernel_size=3, activation="relu"))
model1.add(layers.GlobalMaxPool1D())
model1.add(layers.Dropout(0.5))
model1.add(layers.Dense(8, activation="softmax"))
model1.summary()
model1.compile(optimizer=optimizers.RMSprop(),
def build_model(self):
"""Build an actor (policy) network that maps states -> actions."""
if self.layer_type == "Dense":
# Define input layer (states)
states = layers.Input(shape=(self.state_size, ), name='states')
# Add hidden layers
net = layers.Dense(units=32, activation='relu')(states)
net = layers.Dropout(0.05)(net)
net = layers.Dense(units=64,
activation='relu',
kernel_regularizer=l2(0.01),
activity_regularizer=l1(0.01))(net)
net = layers.Dropout(0.02)(net)
net = layers.Dense(units=32,
activation='relu',
kernel_regularizer=l2(0.01),
activity_regularizer=l1(0.01))(net)
# Try different layer sizes, activations, add batch normalization, regularizers, etc.
elif self.layer_type == "Conv":
# Define input layer (states)
# state_size is 18 (action_repeat[3] * state dim[6])
states = layers.Input(shape=(self.state_size, 4), name='states')
net = layers.Conv1D(filters=32,
kernel_size=2,
padding='same',
activation='relu')(states)
net = layers.MaxPooling1D(pool_size=2)(net)
net = layers.Conv1D(filters=64,
kernel_size=2,
padding='same',
activation='relu')(net)
net = layers.MaxPooling1D(pool_size=2)(net)
net = layers.GlobalAveragePooling1D()(net)
net = layers.Dense(units=128, activation='relu')(net)
else:
raise ValueError(
"Specify layer_type with either 'Dense' or 'Conv'")
# Add final output layer with sigmoid activation
raw_actions = layers.Dense(units=self.action_size,
activation='sigmoid',
name='raw_actions')(net)
# Scale [0, 1] output for each action dimension to proper range
actions = layers.Lambda(lambda x:
(x * self.action_range) + self.action_low,
name='actions')(raw_actions)
# Create Keras model
self.model = models.Model(inputs=states, outputs=actions)
# Define loss function using action value (Q value) gradients
action_gradients = layers.Input(shape=(self.action_size, ))
loss = K.mean(-action_gradients * actions)
# Incorporate any additional losses here (e.g. from regularizers)
# Define optimizer and training function
optimizer = optimizers.Adam()
updates_op = optimizer.get_updates(params=self.model.trainable_weights,
loss=loss)
self.train_fn = K.function(
inputs=[self.model.input, action_gradients,
K.learning_phase()],
outputs=[],
updates=updates_op)
def __init__(self, config, word_vecs, ch_vecs, features, ent_vecs):
""" Initialize all layers. """
self._config = config
self._feature_size = len(features)
self.emb_word = layers.Embedding(len(word_vecs),
300,
input_length=20,
weights=[word_vecs],
trainable=True,
name='word-emb')
self.emb_ent = layers.Embedding(len(ent_vecs),
300,
weights=[ent_vecs],
trainable=True,
name='ent-emb')
# Character CNN over candidate entity title and context
if self._config['character_cnn']:
self.window_size = config['ccnn_window_size']
# universal character embeddings
self.emb_ch = layers.Embedding(len(ch_vecs),
300,
weights=[ch_vecs],
trainable=True,
name='char-emb')
self.ent_reduc = layers.Dense(300, name='ent-reduc-layer')
self.ch_ctx_cnn = layers.Conv1D(100,
self.window_size,
activation='relu',
name='ch-cnn')
self.ch_title_cnn = layers.Conv1D(100,
self.window_size,
activation='relu',
name='ch-title-cnn')
self.ch_ctx_pool = layers.MaxPooling1D(pool_size=MAX_CTX_CHAR_LENGTH
- self.window_size + 1,
name='ch-pool')
self.ch_title_pool = layers.MaxPooling1D(pool_size=MAX_TITLE_CHAR_LNGTH
- self.window_size + 1,
name='ch-titile-pool')
else:
self.ch_ctx_cnn = None
self.ch_ctx_pool = None
self.ch_title_cnn = None
self.ch_title_pool = None
# left and right context encoders w/ attention
self.cell_size = 300
self.left_rnn = layers.GRU(self.cell_size, return_sequences=True, name='left-rnn')
self.right_rnn = layers.GRU(self.cell_size, return_sequences=True, name='right-rnn')
self.left_attn = layers.Dense(self.cell_size, name='left-attn')
self.right_attn = layers.Dense(self.cell_size, name='right-attn')
self.lattn_dist = layers.TimeDistributed(self.left_attn, name='lattn-dist')
self.rattn_dist = layers.TimeDistributed(self.right_attn, name='rattn-tdist')
# binary classification layer
self.reduce_layer = layers.Dense(1, activation='relu', name='final-reduce-layer')
embedding_matrix = create_embedding_matrix("w2vec_wiki_id_300.txt",
tokenizer.word_index, embedding_dim)
nonzero_elements = np.count_nonzero(np.count_nonzero(embedding_matrix, axis=1))
# print(nonzero_elements / vocab_size)
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import RandomizedSearchCV
# CNN with w2vec
cnn_model = Sequential()
cnn_model.add(
layers.Embedding(vocab_size,
embedding_dim,
weights=[embedding_matrix],
input_length=maxlen))
cnn_model.add(layers.Conv1D(200, 5, activation="relu"))
cnn_model.add(layers.GlobalMaxPooling1D())
cnn_model.add(layers.Dense(10, activation="relu"))
cnn_model.add(layers.Dense(1, activation="sigmoid"))
cnn_model.compile(optimizer="adam",
loss="binary_crossentropy",
metrics=["accuracy"])
cnn_model.summary()
history = cnn_model.fit(
X_train,
y_train,
epochs=100,
verbose=False,
validation_data=(X_test, y_test),
batch_size=10,
yAv_train = y_Av[train_idx]
v_train = v[train_idx]
############## DMP Model Design
in_shp_yAv = np.shape(yAv_train)[1:]
in_shp_v = np.shape(v_train)[1:]
### construct neural network graph
n_channels = 32
input_yAv = layers.Input(shape=(10, ))
input_v = layers.Input(shape=(rows, cols))
yAv = layers.Reshape(in_shp_yAv + (1, ))(input_yAv)
v_in = layers.Reshape(in_shp_v + (1, ))(input_v)
for _ in range(3):
yAv = layers.Conv1D(rows,
3,
data_format="channels_first",
activation='relu',
padding='same')(yAv)
for _ in range(3):
yAv = layers.Conv1D(cols,
3,
data_format="channels_last",
activation='relu',
padding='same')(yAv)
yAv2D = layers.Reshape((rows, cols, 1))(yAv)
H_stack = layers.concatenate([yAv2D, v_in], axis=-1)
H = layers.Conv2D(n_channels, (3, 3), activation='relu',
padding='same')(H_stack)
for _ in range(5):
H = layers.Conv2D(n_channels, (3, 3), activation='relu', padding='same')(H)
np_load_old = np.load
np.load = lambda *a, **k: np_load_old(*a, allow_pickle=True, **k)
max_features = 2000
max_len = 500
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
x_train = sequence.pad_sequences(x_train, maxlen=max_len)
x_test = sequence.pad_sequences(x_test, maxlen=max_len)
model = keras.models.Sequential()
model.add(
layers.Embedding(max_features, 128, input_length=max_len, name='embed'))
model.add(layers.Conv1D(32, 7, activation='relu'))
model.add(layers.MaxPooling1D(5))
model.add(layers.Conv1D(32, 7, activation='relu'))
model.add(layers.GlobalMaxPooling1D())
model.add(layers.Dense(1))
model.summary()
model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['acc'])
callbacks = [
keras.callbacks.TensorBoard(log_dir='log_tensor', histogram_freq=1)
]
history = model.fit(x_train,
y_train,
epochs=20,
batch_size=128,
print(input_train)
print(input_train.shape)
input_val=np.zeros((len(X_val),64,3))
input_val[:,:,0]=X_val[:,:64]
input_val[:,:,1]=X_val[:,64:128]
input_val[:,:,2]=X_val[:,128:]
print(input_val)
print(input_val.shape)
model = Sequential()
model.add(layers.Conv1D(512,
1,
activation='relu',
input_shape=(64, 3)))
model.add(Flatten())
model.add(layers.Dense(512,
activation='relu'
)
)
model.add(layers.Dense(128,
activation='relu'
)
)
#The multi-classifical problem,so we use softmax with 46 NN union to achieve the purpose.
x=model.add(layers.Dense(99,
activation='softmax'
